To simplify the point cloud data while preserving features, a novel algorithm based on the curvature distance isput forward. The whole point cloud is divided into a series of initial sub-clusters with the 3D grid subdivision method, and then k neighborhood is constructed from the partition results. All the points in k neighborhood are approximated by quadratic parametric surface based on scattered point cloud parameterization. The curvatures of fitting surface are further calculated. The judgment of requiring reduction is decided by the novel minimal surface distance of curvature features. Some typical cases with various surface features, such as surf, stone, pottery figurine and tooth, are chosen to verify the new method. The results indicate that the new algorithm is of significance in theory and practice for reduction of point cloud, and enables to reduce data directly and efficiently while maintaining the geometry of the original model. The reliability and accuracy of the algorithm are also proved by experiment.